Journal of Computational and Applied Mathematics 260 (2014) 449–469

Contents lists available at ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

An improved method based on Haar wavelets for numerical

solution of nonlinear integral and integro-differential
equations of first and higher orders
Siraj-ul-Islam a , Imran Aziz a,b,∗ , A.S. Al-Fhaid c
a
Department of Basic Sciences, University of Engineering and Technology, Peshawar, Pakistan
b
Department of Mathematics, University of Peshawar, Pakistan
c
Department of Mathematics, King Abdulaziz University, Jeddah, Saudi Arabia

article info abstract

Article history: In this paper, a novel technique is being formulated for the numerical solution of integral
Received 6 February 2013 equations (IEs) as well as integro-differential equations (IDEs) of first and higher orders.
Received in revised form 5 September 2013 The present approach is an improved form of the Haar wavelet methods (Aziz and
Siraj-ul-Islam, 2013, Siraj-ul-Islam et al., 2013). The proposed modifications resulted in
Keywords: computational efficiency and simple applicability of the earlier methods (Aziz and Siraj-
Haar wavelet
ul-Islam, 2013, Siraj-ul-Islam et al., 2013). In addition to this, the new approach is being
Fredholm integral equations
Volterra integral equations
extended from IDEs of first order to IDEs of higher orders with initial- and boundary-
First-order integro-differential equations conditions. Unlike the methods (Aziz and Siraj-ul-Islam, 2013, Siraj-ul-Islam et al., 2013)
Second-order integro-differential equations (where the kernel function is being approximated by two-dimensional Haar wavelet),
Fourth-order integro-differential equations the kernel function in the present case is being approximated by one-dimensional Haar
wavelet. The modified approach is easily extendable to higher order IDEs. Numerical
examples are being included to show the accuracy and efficiency of the new method.
© 2013 Elsevier B.V. All rights reserved.

1. Introduction

Integral equations appear in the mathematical formulations of a variety of modeling procedures. These include jump
diffusion and option pricing, fluid dynamics, biomedical areas (e.g. population biology), chemical kinetics, ecology, control
theory of financial mathematics, aerospace systems, industrial mathematics, etc. Initial- and boundary-value problems in
IDEs have several applications in the study of physical, biological and chemical sciences [1–8]. The numerical solution of IEs
as well as IDEs becomes an imperative as an analytical solution does not exist for complicated IEs and IDEs. A numerical
solution of IEs and of different order IDEs is the subject of recent research. Some recent relevant contributions in this area
include the Haar wavelet method [9,10], the block boundary-value method [11], the Adomian decomposition method [12],
the Hybrid Legendre polynomials and Block–Pulse functions approach [13], the variational iteration method (VIM) [14], the
homotopy perturbation method (HPM) [15], the Sinc collocation method [16], the meshless method [17] and the sequential
approach [18].
Some more relevant work in the area of integral and integro-differential equations include semi-analytical methods
such as the Taylor-series expansion method [19], the meshless collocation method [20], Haar functions method [21,22],
the Legendre multi-wavelets method [23], Nystrom methods [24]. Other numerical methods for solving nonlinear integral

∗ Corresponding author at: Department of Basic Sciences, University of Engineering and Technology, Peshawar, Pakistan. Tel.: +92 3458518182.
E-mail addresses: [email protected], [email protected] (I. Aziz).

0377-0427/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.cam.2013.10.024
450 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

equations can be found in the Refs. [25–28]. An overview of the existing numerical methods establishes the fact that each of
the existing methods whether semi-analytical or fully numerical has limited scope in terms of its applicability to a specific
type of kernel function present in the IEs and IDEs. A great need is being felt to have a single platform for nonlinear IEs
and nonlinear IDEs of various order. To this end, we embarked on modifying our earlier methods [9,10]. The new method
developed in this paper is fundamentally different from the above existing methods [16–18] and Nystrom method [24] in
the following sense:
i The new procedure uses nowhere the existing numerical quadrature for the integration of the kernel function.
ii The new method uses a new procedure for finding coefficients.
iii The new method is more accurate due to exact integration of the Haar basis functions. The distinct feature of the new
procedure is that only the kernel function is being approximated by Haar functions and subsequently the integral is
exactly evaluated by integrating the Haar functions. On the other hand the existing methods [16–18,24] which are based
on Nystrom’s procedure [24] are subject to approximation technique twice, once in the approximation of kernel function
and second in evaluating integration of the kernel function numerically by the existing quadrature formulae. Due to this
fundamental difference, these methods are more prone to errors in the final solution.
iv The new method based on the Haar wavelet is the natural choice when there is a discontinuity either in the kernel
function or in the initial conditions.
v The new method can tackle singularity present in either the derivative part or in the kernel function or at either end
point in the domain interval.
vi The new method can be applied easily and accurately to different types of nonlinear Fredholm and Volterra IEs and IDEs
with initial and boundary conditions.
In the previous work [9,10], the kernel function K has been approximated by two-dimensional Haar wavelet. In the case
of the Haar wavelet method [9] for IEs, the approximation of the kernel function is given as follows:
2M 
 2N
K (x, t , u(t )) ≈ bi,j hi (x)hj (t ), (1)
i=1 j=1

while in case of the Haar wavelet method [10] for IDEs, the approximation is given as:
2M 
 2N
K (x, t , u(t ), u′ (t )) ≈ bi,j hi (x)hj (t ). (2)
i=1 j=1

In both the cases the unknown coefficients bi,j , i, = 1, 2, . . . , 2M , j = 1, 2, . . . , 2N can be expressed as a sum of function
values K (xp , tq , u(tq )) in the case of IEs and function values K (xp , tq , u(tq ), u′ (tq )) in the case of IDEs for certain values
of integers p and q. For large values of M and N, the methods [10,9] involve 4MN number of unknown coefficients and
therefore, become computationally intensive. To enhance computational efficiency of the methods, the two-dimensional
Haar approximations given in Eqs. (1) and (2) are replaced by one-dimensional Haar wavelet basis. The resulting modified
one-dimensional approximation for the kernel function of IEs can be written as:
2M

K (x, t , u(t )) ≈ ai (x)hi (t ). (3)
i =1

Similarly the modified one-dimensional approximation for the kernel function of IDEs of first-order is given as:
2M

K (x, t , u(t ), u′ (t )) ≈ ai (x)hi (t ). (4)
i=1

Due to this fundamental shift, the computational cost of the new algorithm reduces considerably as it involves only 2M
unknown coefficients ai , i = 1, 2, . . . , 2M. The proposed method is employed for the numerical solution of IEs as well
as IDEs having first-, second- and fourth-orders including problems having singular terms in the derivative term at either
end points of the domain interval, having discontinuous kernel function and having non-differentiable/discontinuous exact
solution.
The rest of the paper is organized as follows. In Section 2, we give a brief introduction of Haar wavelet functions and
their integrals. In Section 3, we present a solution procedure of the new method. In Section 4, numerical experiments are
presented for validation of the new algorithm. In Section 5, we give some conclusions of the work.

2. Haar wavelet

A wavelet family (ψj,i (x))j∈N,i∈Z is an orthonormal subfamily of the Hilbert space L2 (R) with the property that all function
in the wavelet family are generated from a fixed function ψ called the mother wavelet through dilations and translations.
The wavelet family satisfies the following relation:
ψj,i (x) = 2j/2 ψ(2j x − i).
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 451

The Haar wavelet family defined on the interval [0, 1) consists of the following functions:
for x ∈ [0, 1)

1
h1 (x) =
0 elsewhere
and
1 for x ∈ [α, β)

hi (x) = −1 for x ∈ [β, γ )
0 elsewhere, i = 2, 3, . . . ,
where
k (k + 0.5) (k + 1)
α= , β= , γ = ;
m m m
m = 2j , j = 0, 1, . . . , k = 0, 1, . . . , m − 1.
The integer j indicates the level of the wavelet and k is the translation parameter. The relation between i, m and k is given
by i = m + k + 1. The function h1 (x) is called the scaling function whereas h2 (x) is the mother wavelet for the Haar wavelet
family.
Any square integrable function f (x) defined on [0, 1) can be expressed as a linear combination of members of Haar
wavelet family and is given as follows:
∞

f ( x) = ai hi (x),
i=1

where ai are constants.

For approximation purposes we consider a maximum value J of the integer j level of the Haar wavelet in the above
definition. The integer J is then called the maximum level of resolution. We also define the integer M = 2J . With these
notations any square integrable function f (x) defined on [0, 1) can be approximated as a linear combination of finite
members of the Haar wavelet family and is given as follows:
2M

f ( x) ≈ ai hi (x). (5)
i=1

The following notations are introduced:

 x
pi,1 (x) = hi (x′ ) dx′ ,
0
 x
pi,n+1 (x) = pi,n (x′ ) dx′ , n = 1, 2, . . . ,
0

and
 1
Ci,n = pi,n (x) dx, n = 1, 2, . . . .
0

These integrals can be evaluated using the definition of Haar wavelet and are given as follows:
for x ∈ [0, α)
0

1

(x − α)n for x ∈ [α, β)




 n!

pi,n (x) = 1 
(x − α)n − 2(x − β)n for x ∈ [β, γ )


 n!



 1 (x − α)n − 2(x − β)n + (x − γ )n 

for x ∈ [γ , 1), n = 1, 2, . . . ,

n!
and
1
(1 − α)n+1 − 2(1 − β)n+1 + (1 − γ )n+1 , n = 1, 2, . . . ,
 
Ci,n =
(n + 1)!
where i = 2, 3, . . .. For i = 1, we have
xn
p1,n (x) = , n = 1, 2, . . . ,
n!
and
1
C1,n = , n = 1, 2, . . . .
(n + 1)!
452 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

3. Solution procedure

In this section we will discuss the proposed numerical methods for two different types (i.e., Fredholm and Volterra) of
IEs as well as IDEs of first- and higher-orders. In the first subsection we state a theorem for efficient evaluation of one-
dimensional Haar wavelet approximation, while in the second and third subsections theoretical results are applied in order
to find the numerical solution of Fredholm and Volterra IEs. In the fourth and fifth subsections the results are extended to
Fredholm and Volterra IDEs. In the sixth and seventh subsections we apply these results for finding numerical solution of
second-order initial-value Fredholm and Volterra IDEs. In the eighth subsection these results are extended to the numerical
solution of second-order boundary-value Fredholm and Volterra IDEs. In the ninth subsection a numerical method is derived
for fourth-order boundary-value Fredholm and Volterra IDEs.
For Haar wavelets approximations the following collocation points are considered:
p − 0.5
xp = , p = 1, 2, . . . , 2M , (6)
2M
q − 0.5
tq = , q = 1, 2, . . . , 2M . (7)
2M

3.1. One-dimensional Haar wavelet system

Consider a square integrable function f (x) defined on the interval [0, 1) and its Haar wavelet approximation given in
Eq. (5). Substituting the collocation points given in Eq. (6) into Eq. (5), we obtain the following linear system of equations:
2M

f (xp ) = ai hi (xp ), p = 1, 2, . . . , 2M . (8)
i =1

The system given in Eq. (8) is a 2M × 2M linear system of equations whose solution for the unknown coefficients ai can be
calculated using the following theorem.

Theorem 1. The solution of the system (8) is given as follows:

2M
1 
a1 = f (xp ),
2M p=1
β γ
 
1  
ai = f (xp ) − f (xp ) , i = 2, 3, . . . , 2M ,
ρ p=α p=β+1

where
α = ρ(σ − 1) + 1,
ρ
β = ρ(σ − 1) + ,
2
γ = ρσ ,
(9)
2M
ρ= ,
τ
σ = i − τ,
τ = 2⌊log2 (i−1)⌋ .

Proof. See [9]. 

Next consider a square integrable function f (x, t ) of two variables. Using the Haar wavelet basis, this function f (x, t ) can
be approximated as follows:
2M

f (x, t ) = ai (x)hi (t ).
i=1

Substituting the collocation points given in Eq. (7), we obtain the following linear system of equations:
2M

f (x, tq ) = ai (x)hi (tq ), q = 1, 2, . . . , 2M . (10)
i=1

The system given in Eq. (10) is a 2M × 2M linear system of equations whose solution for the unknown coefficients ai (x) for
any value of x can be calculated using the following corollary:
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 453

Corollary 1. The solution of the system (10) is given as follows:

2M
1 
a1 (x) = f (x, tq ),
2M q=1
β γ
 
1  
ai (x) = f (x, tq ) − f (x, tq ) , i = 2, 3, . . . , 2M ,
ρ q=α q=β+1

where α, β, γ and ρ are defined as in Eq. (9).

Corollary 2. Suppose that a square integrable function f (x) defined on the interval [0, 1) is approximated using Haar wavelet
expansion given in Eq. (5) and suppose f (x) is known at the collocation points xp , p = 1, 2, . . . , 2M. Then at any point x ∈ [0, 1),
the approximate value of f (x) is given by the following expression:
β γ
 
2M 2M
  1  
f (x) = f (xq )h1 (x) + f (xq ) − f (xq ) hi (x),
q =1 i=2
ρ q=α q=β+1

where α, β, γ and ρ are defined as in Eq. (9).

Corollary 3. Suppose f (x) is a square integrable function defined on the interval [0, 1) and f ′ (x) is approximated using the Haar
wavelet expansion given in Eq. (5). Let us assume that f ′ (x) is known at the collocation points xp , p = 1, 2, . . . , 2M. Then at any
point x ∈ [0, 1), the approximate value of f (x) is given by the following expression:

β γ
 
2M 2M
1  ′  1  
f (x) = f0 + f (xq )p1,1 (x) + f (xq ) −
′
f (xq ) pi,1 (x),
′
(11)
2M q=1 i=2
ρ q=α q=β+1

where f0 = f (0) and α, β, γ , ρ are defined as in Eq. (9).

Proof. Replacing f (x) by f ′ (x) in Corollary 2, we obtain

β γ
 
2M 2M
  1  
f (x) =
′
f (xq )h1 (x) +
′
f (xq ) −
′
f (xq ) hi (x).
′
(12)
q =1 i =2
ρ q=α q=β+1

Integrating Eq. (12) from 0 to x, we obtain Eq. (11). 

Corollary 4. Suppose f (x) is a square integrable function defined on the interval [0, 1) and f ′′ (x) is approximated using the Haar
wavelet expansion given in Eq. (5). Let us assume that f ′′ (x) is known at the collocation points xp , p = 1, 2, . . . , 2M. Then at any
point x ∈ [0, 1), the approximate values of f ′ (x) and f (x) are given by the following expressions:
β γ
 
2M 2M
1  ′′  1  
f (x) = f0 +
′ ′
f (xq )p1,1 (x) + f (xq ) −
′′
f (xq ) pi,1 (x),
′′
2M q=1 i=2
ρ q=α q=β+1

and
β γ
 
2M 2M
1  ′′  1  
f (x) = f0 + xf0 + ′
f (xq )p1,2 (x) + f ( xq ) −
′′
f (xq ) pi,2 (x),
′′
2M q=1 i=2
ρ q=α q=β+1

where f0′ = f ′ (0), f0 = f (0) and α, β, γ , ρ are defined as in Eq. (9).

3.2. Fredholm integral equations

We consider the Fredholm integral equation of the following type:

 1
u(x) = f (x) + K (x, t , u(t )) dt . (13)
0

By using Haar wavelet approximation given in Eq. (3) for the kernel function K (x, t , u(t )), Eq. (13) reduces to the following
form:

u(x) = f (x) + a1 (x).

454 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

Substituting the collocation points given in Eq. (6), we obtain the following system of nonlinear equations:

u(xp ) = f (xp ) + a1 (xp ), p = 1, 2, . . . , 2M .

The value of a1 (xp ) can be replaced with the expression given in Corollary 1 and we obtain the following:

2M
1 
u(xp ) = f (xp ) + K (xp , tq , u(tq )), p = 1, 2, . . . , 2M . (14)
2M q=1

The system given in Eq. (14) is a 2M × 2M system of nonlinear equations which can be solved either using Newton’s method
or Broyden’s method. The solution of the system given in Eq. (14) gives us the approximate values of u(x) at collocation
points xp , p = 1, 2, . . . , 2M. The approximate values of u(x) at any point of the domain other than the collocation points
can be obtained using Corollary 2 and is given by the following relation:

β γ
 
2M 2M
  1  
u(x) = u(xq )h1 (x) + u(xq ) − u(xq ) hi (x). (15)
q=1 i =2
ρ q=α q=β+1

The Jacobian of the system (14) is given as follows:

1 − φ1 (p, p),

p=r
J (p, r ) =
−φ1 (p, r ), p ̸= r

where

1  ∂ K (xp , tq , u(tq )) 1 ∂K
2M
φ1 (p, r ) = = (xp , tr , u(tr )).
2M q=1 ∂ u( x r ) 2M ∂ u

3.3. Volterra integral equations

The following type of Volterra integral equation is considered:

 x
u(x) = f (x) + K (x, t , u(t )) dt . (16)
0

As in the case of Fredholm IEs, the approximation given in Eq. (3) is used for the kernel function K (x, t , u(t )) and after
simplifications we obtain the following equation:
2M

u(x) = f (x) + ai (x)pi,1 (x).
i=1

Substituting the collocation points given in Eq. (6) we get the following:
2M

u(xp ) = f (xp ) + ai (xp )pi,1 (xp ), p = 1, 2, . . . , 2M .
i=1

The values ai (x) can be replaced using Corollary 1 and the following system of equations is obtained:
2M
1 
u(xp ) = f (xp ) + K (xp , tq , u(tq ))p1,1 (xp )
2M q=1

β γ
 
2M
 1  
+ K (xp , tq , u(tq )) − K (xp , tq , u(tq )) pi,1 (xp ) p = 1, 2, . . . , 2M , (17)
i=2
ρ q=α q=β+1

where α, β, γ and ρ are defined as in Eq. (9). Now Eq. (17) represents a 2M × 2M nonlinear system of equations with
unknowns u(xp ), p = 1, 2, . . . , 2M. As in the case of Fredholm IEs we can solve the system (17) using Newton’s or Broyden’s
method and obtain a numerical solution of Eq. (16) at collocation points. The approximate solution at any point of the domain
can then be obtained using the relation given in Eq. (15).
The Jacobian of the system (17) is given below:

1 − φ2 (p, p)

p=r
J (p, r ) =
−φ2 (p, r ) p ̸= r
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 455

where
1 ∂K  ∂K
2M
 1
φ2 (p, r ) = (xp , tr , u(tr )) + I[α,β] (r ) − I[β+1,γ ] (r ) (xp , tr , u(tr ))pi,1 (xp ).
2M ∂ u i=2
ρ ∂u
I denotes the characteristic function defined as follows:

1 if a ≤ x ≤ b
I[a,b] (x) =
0 otherwise.

3.4. First-order Fredholm integro-differential equation

Consider the following type of nonlinear Fredholm integro-differential equation of first-order:

 1
u′ (x) + g (x)u(x) = f (x) + K (x, t , u(t ), u′ (t )) dt , u(0) = u0 . (18)
0
In this case, for the kernel function K (x, t , u(t )), the Haar wavelet approximation given in Eq. (4) is used and with this
approximation Eq. (18) reduces to the following form:
u′ (x) + g (x)u(x) = f (x) + a1 (x).
Substituting the collocation points given in Eq. (6), we obtain the following system of nonlinear equations:
u′ (xp ) + g (xp )u(xp ) = f (xp ) + a1 (xp ), p = 1, 2, . . . , 2M . (19)
The value of a1 (xp ), p = 1, 2, . . . , 2M is substituted from Corollary 1 and thus the system (19) can be written as:
2M
1 
u′ (xp ) + g (xp )u(xp ) = f (xp ) + K (xp , tq , u(tq ), u′ (tq )), p = 1, 2, . . . , 2M . (20)
2M q=1

The values u(xp ), p = 1, 2, . . . , 2M can be expressed in terms of u′ (xp ), p = 1, 2, . . . , 2M using Corollary 3 and is given
as follows:
β γ
 
2M 2M
1  ′  1  
u( x p ) = u0 + u (xq )p1,1 (xp ) + u (xq ) −
′
u (xq ) pi,1 (xp ),
′
p = 1, 2, . . . , 2M , (21)
2M q=1 i=2
ρ q=α q=β+1

where α, β, γ and ρ are defined as in Eq. (9). Thus we conclude that the system given in Eq. (20) contains only the following
unknowns:
u′ (xp ), p = 1, 2, . . . , 2M .
Hence the system given in Eq. (20) is a 2M × 2M nonlinear system which, as in previous cases, can be solved using either
Newton’s method or Broyden’s method. The solution of this system yields the values of u′ (x) at the collocation points
xp , p = 1, 2, . . . , 2M. The approximate solution at any point of the domain can be calculated using Corollary 3 and is given
by the following relation:
β γ
 
2M 2M
  1  
u( x ) = u0 + u (xq )p1,1 (x) +
′
u (xq ) −
′
u (xq ) pi,1 (x),
′

q =1 i=2
ρ q=α q=β+1

where α, β, γ and ρ are defined as in Eq. (9).

The Jacobian of the system (20) is given as follows:
1 + φ1 (p, p),

p=r
J (p, r ) =
φ1 (p, r ), p ̸= r
where
∂ u(xp ) 1  ∂ K (xp , tq , u(tq ), u′ (tq ))
2M
φ1 (p, r ) = g (xp ) − ,
∂ u′ (xr ) 2M q=1 ∂ u′ ( x r )
and
 ∂K ∂ u( x p ) ∂K

∂ K (xp , tq , u(tq ), u (tq ))  ∂ u (xp , tq , u(tq ), u (tq )) ∂ u′ (xr ) + ∂ u′ (xp , tr , u(tr ), u (tr )),
′ ′
′ q=r
= ∂K ∂ u( x p )
∂ u′ (xr )  (xp , tq , u(tq ), u′ (tq ))
 , q ̸= r .
∂u ∂ u′ (xr )
∂ u(x )
The value of partial derivative ∂ u′ (xp ) can be calculated using Eq. (21) and is given as follows:
r

∂ u(xp ) 1
2M
 1
p1,1 (xp ) + I[α,β] (r ) − I[β+1,γ ] (r ) pi,1 (xp ), p, r = 1, 2, . . . , 2M .

=
∂ u (xr )
′ 2M i =2
ρ
456 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

3.5. First-order Volterra integro-differential equation

In this section we will consider the following type of nonlinear Volterra integro-differential equation of first-order:
 x
u′ (x) + g (x)u(x) = f (x) + K (x, t , u(t ), u′ (t )) dt , u(0) = u0 . (22)
0

Substituting the Haar expression for the kernel function K (x, t , u(t ), u′ (t )) given in Eq. (4) and simplifying we obtain the
following equation:
2M

u′ (x) + g (x)u(x) = f (x) + ai (x)pi,1 (x).
i =1

Substituting the collocation points given in Eq. (6) we get the following nonlinear system of equations:
2M

u′ (xp ) + g (xp )u(xp ) = f (xp ) + ai (xp )pi,1 (xp ), p = 1, 2, . . . , 2M .
i=1

Substituting the values of ai (x), i = 1, 2, . . . , 2M using Corollary 1, we obtain the following:

2M
1 
u′ (xp ) + g (xp )u(xp ) = f (xp ) + K (xp , tq , u(tq ), u′ (tq ))p1,1 (xp )
2M q=1
β γ
 
2M
 1  
+ K (xp , tq , u(tq ), u (tq )) −
′
K (xp , tq , u(tq ), u (tq )) pi,1 (xp )
′

i=2
ρ q=α q=β+1

p = 1, 2, . . . , 2M , (23)
where α, β, γ and ρ are defined as in Eq. (9). Now Eq. (23) represent a 2M × 2M system with unknowns u′ (xp ), p =
1, 2, . . . , 2M. As in the previous cases we can solve this system using Newton’s or Broyden’s method and obtain numerical
solution of Eq. (22) at any point of the domain.
The Jacobian of the system (23) is given below.
1 + φ2 (p, p)

p=r
J (p, r ) =
φ2 (p, r ) p ̸= r ,
where
∂ u(xp ) 1  ∂ K (xp , tq , u(tq ), u′ (tq ))
2M
φ2 (p, r ) = g (xp ) −
∂ u (xr ) 2M q=1
′ ∂ u′ (xp )
γ
 β 
1  ∂ K (xp , tq , u(tq ), u′ (tq )) ∂ K (xp , tq , u(tq ), u′ (tq ))
2M
 
− − pi,1 (xp ).
i=2
ρ q=α ∂ u′ (xp ) q=β+1
∂ u′ (xp )

3.6. Second-order initial-value Fredholm integro-differential equation

Consider the following initial-value type of Fredholm integro-differential equation of second-order:

 1
u (x) + g1 (x)u (x) + g (x)u(x) = f (x) +
′′ ′
K (x, t , u(t ), u′ (t ), u′′ (t )) dt , (24)
0

subject to the initial conditions u(0) = u0 , u′ (0) = u′0 . We assume as in the previous cases that the kernel function can be
approximated using the Haar wavelet series as follows:
2M

K (x, t , u(t ), u′ (t ), u′′ (t )) = ai (x)hi (t ). (25)
i =1

Substituting the above expression and collocation points in Eq. (24), we obtain the following system of nonlinear equations:
u′′ (xp ) + g1 (xp )u′ (xp ) + g (xp )u(xp ) = f (xp ) + a1 (xp ), p = 1, 2, . . . , 2M .
Using Corollary 1 the above system becomes
2M
1 
u′′ (xp ) + g1 (xp )u′ (xp ) + g (xp )u(xp ) = f (xp ) + K (xp , tq , u(tq ), u′ (tq ), u′′ (tq )), p = 1, 2, . . . , 2M . (26)
2M q=1
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 457

By using Corollary 4 we can express u′ (xp ), p = 1, 2, . . . , 2M and u(xp ), p = 1, 2, . . . , 2M in terms of u′′ (xp ), p =
1, 2, . . . , 2M and are given below:

γ
 β 
2M 2M
1  ′′  1  
u (xp ) = u0 +
′ ′
u (xq )p1,1 (xp ) + u (xq ) −
′′
u (xq ) pi,1 (xp ),
′′
p = 1, 2, . . . , 2M , (27)
2M q=1 i =2
ρ q=α q=β+1

and
2M
1  ′′
u(xp ) = u0 + u′0 xp + u (xq )p1,2 (xp )
2M q=1

β γ
 
2M
 1  
+ u (xq ) −
′′
u (xq ) pi,2 (xp ),
′′
p = 1, 2, . . . , 2M . (28)
i=2
ρ q=α q=β+1

Thus Eq. (26) represents a 2M × 2M nonlinear system with unknowns u′′ (xp ), p = 1, 2, . . . , 2M. Solving this system we
obtain values of u′′ (x) at the collocation points. The numerical solution of the Eq. (24) at any point can be calculated using
Corollary 4 and is given as follows:
γ
 β 
2M 2M
1  ′′  1  
u(x) = u0 + u0 x + ′
u (xq )p1,2 (x) + u (xq ) −
′′
u (xq ) pi,2 (x).
′′
2M q=1 i =2
ρ q=α q=β+1

The Jacobian of the system (26) is given as follows:

1 + φ3 (p, p)

p=r
J (p, r ) =
φ3 (p, r ) p ̸= r

where

∂ u′ (xp ) ∂ u(xp ) 1  ∂ K (xp , tq , u(tq ), u′ (tq ), u′′ (tq ))

2M
φ3 (p, r ) = g1 (xp ) + g (xp ) ′′ − ,
∂ u (xr )
′′ ∂ u (xr ) 2M q=1 ∂ u′′ (xp )

and
∂ K (xp , tq , u(tq ), u′ (tq ), u′′ (tq ))
∂ u′′ (xr )
∂K ∂ u(xp ) ∂K ∂ u′ (xp )

 (xp , tq , u(tq ), u′ (tq ), u′′ (tq )) ′′ + (xp , tq , u(tq ), u′ (tq ), u′′ (tq )) ′′
∂ ∂ u (xr ) ∂u ∂ u (xr )


 u ′
∂K


= + ′′ (xp , tr , u(tr ), u′ (tr ), u′′ (tq )), q=r
∂ u
∂K ∂ u(xp ) ∂K ∂ u′ (xp )



 (xp , tq , u(tq ), u′ (tq ), u′′ (tq )) ( , , ( ), ′
( ), ′′
( )) , q ̸= r .

 + x p tq u tq u t q u t q
∂u ∂ u′′ (xr ) ∂ u′ ∂ u′′ (xr )
∂ u′ (x ) ∂ u(x )
The values of the partial derivatives ∂ u′′ (xp ) and ∂ u′′ (xp ) can be calculated using Eqs. (27) and (28) respectively and are given
r r
as follows:

∂ u′ (xp ) 1
2M
 1
( ) I[α,β] (r ) − I[β+1,γ ] (r ) pi,1 (xp ), p, r = 1, 2, . . . , 2M ,

= p 1, 1 x p +
∂ u′′ (xr ) 2M i=2
ρ
and

∂ u(xp ) 1
2M
 1
p1,2 (xp ) + I[α,β] (r ) − I[β+1,γ ] (r ) pi,2 (xp ), p, r = 1, 2, . . . , 2M .

=
∂ u ( xr )
′′ 2M i=2
ρ

3.7. Second-order initial-value Volterra integro-differential equation

We consider the following initial-value type of nonlinear Volterra IDE of second-order:

 x
u′′ (x) + g1 (x)u′ (x) + g (x)u(x) = f (x) + K (x, t , u(t ), u′ (t ), u′′ (t )) dt , (29)
0
458 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

with initial conditions u(0) = u0 , u′ (0) = u′0 . Similar to previous cases, substituting the Haar wavelet expression given in
Eq. (25), collocation points and using the Corollary 1, the following system of nonlinear equations is obtained:
2M
1 
u′′ (xp ) + g1 (xp )u′ (xp ) + g (xp )u(xp ) = f (xp ) + K (xp , tq , u(tq ), u′ (tq ), u′′ (tq ))p1,1 (xp )
2M q=1
β γ
 
2M
 1  
+ K (xp , tq , u(tq ), u (tq ), u (tq )) −
′ ′′
K (xp , tq , u(tq ), u (tq ), u (tq )) pi,1 (xp )
′ ′′

i=2
ρ q=α q=β+1

p = 1, 2, . . . , 2M . (30)
The procedure for finding numerical solution in this case is similar to the case of Fredholm IDE of the second order as
discussed in the previous section.
The Jacobian of the system (30) is given below:
1 + φ4 (p, p)

p=r
J (p, r ) =
φ4 (p, r ) p ̸= r ,
where
∂ u′ (xp ) ∂ u(xp ) 1  ∂ K (xp , tq , u(tq ), u′ (tq ), u′′ (tq ))
2M
φ4 (p, r ) = g1 (xp ) + g (x p ) −
∂ u′′ (xr ) ∂ u′′ (xr ) 2M q=1 ∂ u′′ (xp )
γ
 β 
1  ∂ K (xp , tq , u(tq ), u′ (tq ), u′′ (tq )) ∂ K (xp , tq , u(tq ), u′ (tq ), u′′ (tq ))
2M
 
− − pi,1 (xp ).
i=2
ρ q=α ∂ u′′ (xp ) q=β+1
∂ u′′ (xp )

3.8. Second order boundary-value Fredholm and Volterra integro-differential equation

In this section we consider the nonlinear boundary-value Fredholm IDE given in Eq. (24) and nonlinear Volterra IDE
given in Eq. (29) with boundary conditions u(0) = u0 , u(1) = u1 . The procedures for finding the numerical solution in
these cases are similar to the cases with initial conditions, respectively, except we have different expressions for u′ (xp ), p =
1, 2, . . . , 2M and u(xp ), p = 1, 2, . . . , 2M in terms of u′′ (xp ), p = 1, 2, . . . , 2M. These expressions are given below. Further
details about calculating these expressions can be found in [29].
2M
1  ′′
u′ (xp ) = u1 − u0 + u (xq )(p1,1 (xp ) − C1,1 )
2M q=1
β γ
 
2M
 1  
+ u (xq ) −
′′
u (xq ) (pi,1 (xp ) − Ci,1 ),
′′
p = 1, 2, . . . , 2M ,
i =2
ρ q=α q=β+1

and
2M
1  ′′
u(xp ) = u0 + (u1 − u0 )xp + u (xq )(p1,2 (xp ) − xp C1,1 )
2M q=1
β γ
 
2M
 1  
+ u ( xq ) −
′′
u (xq ) (pi,2 (xp ) − xp Ci,1 ),
′′
p = 1, 2, . . . , 2M .
i=2
ρ q=α q=β+1

3.9. Fourth-order boundary-value Fredholm and Volterra integro-differential equation

We will consider the following nonlinear boundary-value Fredholm IDE of fourth-order:

 1
(iv)
u (x) + g3 (x)u (x) + g2 (x)u (x) + g1 (x)u (x) + g (x)u(x)‘ = f (x) +
′′′ ′′ ′
K (x, t , u(t ), u′ (t ), u′′ (t ), u′′′ (t ), u(iv) (t ))dt ,
0

and the following nonlinear boundary-value Volterra IDE of fourth-order:

 x
u(iv) (x) + g3 (x)u′′′ (x) + g2 (x)u′′ (x) + g1 (x)u′ (x) + g (x)u(x) = f (x) + K (x, t , u(t ), u′ (t ), u′′ (t ), u′′′ (t ), u(iv) (t ))dt ,
0

subject to the following boundary conditions:

u(0) = u0 , u(1) = u1 , u′ (0) = u′0 , u′ (1) = u′1 .
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 459

Using these boundary conditions and Corollary 1, we can express u(xp ), u′ (xp ), u′′ (xp ), u′′′ (xp ), p = 1, 2, . . . , 2M in terms
of u(iv) (xp ), p = 1, 2, . . . , 2M as follows:
2M
1  (iv)
u′′′ (xp ) = 12u0 − 12u1 + 6u′0 + 6u′1 + u (xq )(p1,1 (xp ) − 6C1,2 + 12C1,3 )
2M q=1
β γ
 
2M
 1 
(iv)

(iv)
+ u (xq ) − u (xq ) (pi,1 (xp ) − 6Ci,2 + 12Ci,3 ), p = 1, 2, . . . , 2M ,
i=2
ρ q=α q=β+1

u′′ (xp ) = −6u0 + 6u1 − 4u′0 − 2u′1 + (12u0 − 12u1 + 6u′0 + 6u′1 )xp
2M
1  (iv)
+ u (xq )(p1,2 (xp ) + (−6C1,2 + 12C1,3 )xp + 2C1,2 − 6C1,3 )
2M q=1
β γ
 
2M
 1 
(iv)

(iv)
+ u (xq ) − u (xq ) (pi,2 (xp ) + (−6Ci,2 + 12Ci,3 )xp + 2Ci,2 − 6Ci,3 ),
i =2
ρ q=α q=β+1
p = 1, 2, . . . , 2M ,
x2p
u′ (xp ) = u′0 + (−6u0 + 6u1 − 4u′0 − 2u′1 )xp + (12u0 − 12u1 + 6u′0 + 6u′1 )
 2 
2M
1  (iv) x2p
+ u (xq ) p1,3 (xp ) + (−6C1,2 + 12C1,3 ) + (2C1,2 − 6C1,3 )xp
2M q=1 2
β γ
  
2M
 1 
(iv)

(iv)
x2p
+ u (xq ) − u (xq ) pi,3 (xp ) + (−6Ci,2 + 12Ci,3 ) + (2Ci,2 − 6Ci,3 )xp ,
i =2
ρ q=α q=β+1
2
p = 1, 2, . . . , 2M ,
and
x2p x3p
u(xp ) = u0 + u′0 xp + (−6u0 + 6u1 − 4u′0 − 2u′1 ) + (12u0 − 12u1 + 6u′0 + 6u′1 )
2
 6
2M
1  (iv) x3p x2p
+ u (xq ) p1,4 (xp ) + (−6C1,2 + 12C1,3 ) + (2C1,2 − 6C1,3 )
2M q=1 6 2
β γ
  
2M
1 x3p x2p
u(iv) (xq ) − u(iv) (xq )
  
+ pi,4 (xp ) + (−6Ci,2 + 12Ci,3 ) + (2Ci,2 − 6Ci,3 ) ,
i=2
ρ q=α q=β+1
6 2
p = 1, 2, . . . , 2M .
The rest of the procedure is similar to previous cases.

4. Numerical experiments

This section is devoted to numerical validation of the proposed theoretical work performed in the previous sections.
The test problems include: nonlinear Fredholm and Volterra IEs [9] (test problems 1 and 2), nonlinear Fredholm IE with
discontinuous kernel function (test problem 3), nonlinear Fredholm IE with non-differentiable exact solution (test problem
4), nonlinear Fredholm IE with discontinuous exact solution (test problem 5 and 6), nonlinear Fredholm IDEs [10,17] (test
problems 7 and 8), nonlinear first-order Volterra–Fredholm IDE [13,30] (test problem 9), nonlinear second-order Fredholm
IDE [31] (test problem 10), boundary-value problem of nonlinear second-order Volterra IDE [32] type (test problem 11),
nonlinear second-order Volterra–Fredholm IDE having singularity at 0 [33] (test problem 12), nonlinear fourth-order
Fredholm IDE (test problem 13), nonlinear fourth-order Volterra IDE [34] (test problem 14).
The experimental rate of convergence Rc (N ) for some selected problems is also calculated which is defined as:
log[Ec (N /2)/Ec (N )]
Rc (N ) = . (31)
log 2
In all test problems we have used Broyden’s method. The initial guess for Broyden’s method was taken to be zero and the
iterations were terminated when the convergent criterion 10−10 was satisfied.

Test Problem 1. Consider the following nonlinear Fredholm IE [9]:

x √   1 
u(x) = −x2 − 2 2−1 +2+ xt u(t )dt .
3 0
460 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

Table 1
Comparison of CPU time of present method with Haar wavelet method [9] for
Test Problem 1.
2M L∞ CPU time CPU time
(Present method) (Haar wavelet method [9])

2 1.22 × 10−2 0.001 0.002

4 3.70 × 10−3 0.004 0.004
8 1.00 × 10−3 0.005 0.03
16 2.60 × 10−4 0.007 0.1
32 6.61 × 10−5 0.01 0.7
64 1.66 × 10−5 0.02 5
128 4.18 × 10−6 0.04 43
256 1.05 × 10−6 0.1 346
512 2.62 × 10−7 0.7 –
1024 6.55 × 10−8 3 –
2048 1.63 × 10−8 12 –
4096 4.02 × 10−9 57 –
8192 1.02 × 10−9 324 –

Table 2
Comparison of CPU time of present method with Haar wavelet method [9] for
Test Problem 2.
2M L∞ CPU time CPU time
(Present method) (Haar wavelet method [9])

2 1.43 × 10−2 0.002 0.01

4 6.28 × 10−3 0.006 0.05
8 2.02 × 10−3 0.007 0.3
16 5.75 × 10−4 0.01 4
32 1.53 × 10−4 0.01 50
64 3.97 × 10−5 0.02 727
128 1.01 × 10−5 0.02 –
256 2.54 × 10−6 0.09 –
512 6.38 × 10−7 0.5 –
1024 1.60 × 10−7 3 –
2048 4.00 × 10−8 19 –
4096 1.00 × 10−8 133 –
8192 2.50 × 10−9 967 –

The exact solution of this problem is u(x) = 2 − x2 . In Table 1 a comparison of the CPU time of the present method with the
Haar wavelet method [9] is given. From this table it is clear that the present modified method works with amazing efficiency
in comparison with our earlier approach [9]. The earlier approach takes 727 s for 2M = 128 whereas the new approach takes
0.02 s which roughly means that the new algorithm is 25,000 times faster than the previous one [9] in terms of CPU time.
The new approach has been tested to get a solution at a much finer resolution, i.e. 2M = 8192. Even in this case the system
has been solved and good accuracy has been obtained.

Test Problem 2. Consider the following nonlinear Volterra IE [9]:

 x
u(x) = f (x) + xt 2 (u(t ))2 dt ,
0

where
 
11 2 1 2 1 11 5 2
f (x) = 1+ x+ 2
x − 3
x + 4
x ln(x + 1) − (x + x4 )(ln(x + 1))2 − x2 + x3 − x4 .
9 3 3 9 3 9 18 27
The exact solution of the problem is u(x) = ln(x + 1). In Table 2 a comparison of the CPU time of the present method with the
Haar wavelet method [9] is given. As evident from the table, superior performance in terms of computational efficiency is
obtained. The new approach has been tested when the matrix size is 8192 × 8192 in the finer resolution space and accuracy
of the order 10−9 is obtained in this case as well.

Test Problem 3 (Discontinuous Kernel). Consider the following nonlinear IE with a discontinuous kernel:
 1
u(x) = K (x, t , u(t ))dt + f (x),
0
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 461

Table 3
Maximum absolute errors and rate of
convergence for Test Problem 3.
2M L∞ Rc (2M )

2 7.3 × 10 −3

4 4.3 × 10−3 0.76

8 1.1 × 10−3 1.96
16 2.72 × 10−4 2.02
32 6.83 × 10−5 1.99
64 1.71 × 10−5 2.00
128 4.27 × 10−6 2.00
256 1.07 × 10−6 2.00
512 2.67 × 10−7 2.00
1024 6.68 × 10−8 2.00

Table 4
Maximum absolute errors and rate of
convergence for Test Problem 4.
2M L∞ R c (N )

2 1.38 × 10 −3

4 4.07 × 10−4 1.76

8 1.09 × 10−4 1.90
16 2.83 × 10−5 1.95
32 7.18 × 10−6 1.98
64 1.81 × 10−6 1.99
128 4.54 × 10−7 2.00
256 1.14 × 10−7 1.99
512 2.85 × 10−8 2.00
1024 7.12 × 10−9 2.00

where
x ( 1 − t ) 2 u2 x ≤ t,

K (x, t , u(t )) =
t (1 − x)2 u2 x > t,
and
1
f ( x) = −π (x − 1)2 (3 + 2π 2 x(2 + x)) − 3π(x2 − 1) cos(2π x) + 24π 3 sin(π x) + 3x sin(2π x) .
 
24π 3
The exact solution of the problem is u(x) = sin(π x). This test problem ensures that the proposed method can be applied
to problems with discontinuous kernels without making any structural changes in the method. In Table 3 we have shown
the maximum absolute errors for different numbers of collocation points. The maximum absolute error is of order 10−8 for
1024 collocation points and it is obvious that even better accuracy can be obtained by increasing the number of collocation
points. The numerical order of convergence given in the last column of the Table 3 shows that the new method is second
order convergent.

Test Problem 4 (Non-differentiable Exact Solution). Consider the following nonlinear Fredholm IE with a non-differentiable
exact solution:
 1
u(x) = xt sin(u(t ))dt − x + |x − 0.5| + x cos(0.5).
0

The exact solution of the problem is u(x) = |x − 0.5| which is not differentiable at x = 0.5. The method has been applied
to the problem without any modification and accuracy of the order 10−9 has been obtained as shown in Table 4. Like the
previous test problem the numerical order of convergence of the proposed method is almost 2.

Test Problem 5 (Non-differentiable Exact Solution at Both Ends). In this test problem we consider a nonlinear Fredholm IE
whose exact solution is not differentiable at both the end points x = 0, 1.
 1
u(x) = sin(x + t )u(t )2 dt + f (x), (32)
0
√ √
where f (x) = (1 − x)x + 2 cos 21 − 2 sin 21 sin 21 + x . The exact solution of the problem is u(x) = (1 − x)x.
      
Comparison of exact and approximate solutions for different number of collocation points 2M = 16, 32, . . . , 512 is depicted
in Fig. 1. From the figure it is obvious that even for a smaller number of collocation points the exact and approximate solutions
462 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

Fig. 1. Comparison of exact and approximate solutions for different numbers of collocation points for Test Problem 5.

match closely. We have also shown the maximum absolute errors for different numbers of collocation points in Table 5 in
order to show how closely the exact and approximate solutions are related.

Test Problem 6 (Discontinuous Exact Solution). Consider the following nonlinear IE with discontinuous exact solution:
 1
u(x) = xt sin(u(t ))dt + f (x),
0
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 463

Table 5
Maximum absolute errors for
Test Problem 5.
2M L∞

2 7.35 × 10−2
4 1.35 × 10−2
8 3.17 × 10−3
16 7.80 × 10−4
32 1.94 × 10−4
64 4.84 × 10−5
128 1.21 × 10−5
256 3.02 × 10−6
512 7.56 × 10−7
1024 1.89 × 10−7

Table 6
Maximum absolute errors for
Test Problem 6.
2M L∞

2 9.25 × 10−3
4 2.71 × 10−3
8 7.27 × 10−4
16 1.88 × 10−4
32 4.77 × 10−5
64 1.20 × 10−5
128 3.02 × 10−6
256 7.56 × 10−7
512 1.89 × 10−7
1024 4.73 × 10−8

Fig. 2. Comparison of exact solution versus approximate solution for Test Problem 6.

where
     2  
11 1 1 1
x x − 10 cos(1) + √2 cos √2 − sin 8 + 6 sin(1) − 9 sin √2 x ≤ 0.5,



f ( x) =   2 
√
  
11 1 1 1
 x + x −10 cos(1) + √ cos √ + 6 sin(1) − 9 sin √ x > 0.5.

− sin


2 2 8 2

The exact solution of the problem is given by

x√2 x ≤ 0.5,

u(x) =
x x > 0.5.

Maximum absolute errors for different numbers of collocation points are given in Table 6. In Fig. 2 we have shown the
comparison of exact solution with approximate solution for 64 collocation points.
464 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

Table 7
Comparison of CPU time of the present method with the Haar wavelet method
[10] for Test Problem 7.
2M L∞ CPU time CPU time
(Present method) (Haar wavelet method [10])

2 3.60 × 10−2 0.002 0.004

4 1.03 × 10−2 0.003 0.01
8 2.70 × 10−3 0.01 0.04
16 6.95 × 10−4 0.02 0.14
32 1.76 × 10−4 0.03 0.9
64 4.41 × 10−5 0.04 6
128 1.11 × 10−5 0.05 45
256 2.76 × 10−6 0.1 –
512 6.91 × 10−7 0.5 –
1024 1.73 × 10−7 2.5 –
2048 4.32 × 10−8 10 –
4096 1.08 × 10−8 51 –
8192 2.70 × 10−9 298 –

Table 8
Comparison of maximum absolute errors of the present method with the
meshless method [17] for Test Problem 8.
Present Method Meshless Method [17]
N L∞ CPU time N L∞ CPU time

4 9.01 × 10−4 0.007 5 1.72 × 10−3 0.35

8 2.79 × 10−4 0.008 9 2.11 × 10−4 0.39
16 7.85 × 10−5 0.01 17 1.83 × 10−4 0.42
32 2.09 × 10−5 0.02 33 4.10 × 10−5 0.51
64 5.38 × 10−6 0.03 65 2.63 × 10−5 0.68
128 1.37 × 10−6 0.04 129 6.48 × 10−6 1.42
256 3.44 × 10−7 0.06 –
512 8.64 × 10−8 0.2 –
1024 2.16 × 10−8 1 –
2048 5.41 × 10−9 5 –
4096 1.35 × 10−9 32 –
8192 3.39 × 10−10 221 –

Test Problem 7. Consider the nonlinear Fredholm IDE [10]:

 1
1
u′ (x) = − 2(−3 − 3 log2 (2) + log3 (2) + log(64)) + u3 (t )dt , u(0) = 0, 0 < x < 1.
1+x 0

The exact solution of the problem is log(x + 1). Comparison of CPU time of present method with Haar wavelet method [10]
is shown in Table 7. It is clear from the table that the CPU time of the present method reduces drastically in comparison with
the method [10]. For 2M = 128, CPU time of the present method is 0.05 s whereas the method using the two-dimensional
Haar wavelet takes 45 s of CPU time before reaching the final solution. The new algorithm is roughly 1000 times faster than
the method [10].

Test Problem 8. Consider the following first-order nonlinear Fredholm IDE [17]:
 1
1
u (x) = 1 −
′ 3
x + x3 u(t )2 dt , u(0) = 0, 0 < x < 1.
3 0

The exact solution of the problem is u(x) = x. In Table 8, we have shown a comparison of maximum absolute errors and
CPU time with the meshless method [17]. Like in the previous problem, the accuracy of the present method is better than
in the meshless method. It is not only the better accuracy of the new method but reduced CPU time which gives an edge to
the new method based on Haar wavelets over the meshless method.

Test Problem 9. Consider the following Volterra–Fredholm IDE [13,30]:

 1  x
1 1
u′ (x) + u(x) = f (x) + tu3 (t )dt − xu2 (t )dt , u(0) = 0, 0 < x < 1,
4 0 2 0

where
1 1
f (x) = x6 + x2 + 2x − .
10 32
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 465

Table 9
Comparison of pointwise absolute errors of present method with TF method [30] for Test Problem 9.
x Present Method TF method [30] HLBPF [13] Present Method TF method [30] HLBPF [13]
(N = 16) (N = 16) (N = 16) (N = 32) (N = 32) (N = 32)
0 0 0 0 0 0 0
0.1 8.2 × 10−4 9.7 × 10−4 9.2 × 10−4 1.2 × 10−4 1.7 × 10−4 2.6 × 10−4
0.2 4.0 × 10−4 7.0 × 10−4 1.7 × 10−3 1.7 × 10−4 2.5 × 10−4 4.9 × 10−4
0.3 3.0 × 10−4 7.4 × 10−4 2.4 × 10−3 1.5 × 10−4 2.6 × 10−4 7.0 × 10−4
0.4 5.3 × 10−4 1.1 × 10−4 2.9 × 10−3 5.4 × 10−5 1.9 × 10−4 8.7 × 10−4
0.5 4.9 × 10−4 1.6 × 10−4 3.4 × 10−3 1.2 × 10−4 4.1 × 10−5 1.0 × 10−3
0.6 3.7 × 10−4 1.1 × 10−3 4.2 × 10−3 1.4 × 10−5 2.0 × 10−4 1.1 × 10−3
0.7 2.2 × 10−5 8.2 × 10−4 3.8 × 10−3 7.3 × 10−5 2.8 × 10−4 9.7 × 10−4
0.8 1.0 × 10−4 8.2 × 10−4 2.4 × 10−3 5.2 × 10−5 2.8 × 10−4 8.3 × 10−4
0.9 1.2 × 10−4 1.1 × 10−3 3.4 × 10−4 4.8 × 10−5 2.0 × 10−4 1.8 × 10−4
1.0 9.1 × 10−4 1.5 × 10−4 1.5 × 10−3 2.2 × 10−4 3.7 × 10−5 3.4 × 10−4

Table 10
Maximum absolute errors for
Test Problem 10.
2M L∞

2 1.32 × 10−1
4 2.89 × 10−2
8 7.76 × 10−3
16 2.04 × 10−3
32 5.23 × 10−4
64 1.33 × 10−4
128 3.34 × 10−5
256 8.37 × 10−6

The exact solution of the problem is u(x) = x2 . Table 9 shows a comparison of pointwise absolute errors of the present
method with the triangular factorization (TF) method [30] as well as the Hybrid Legendre polynomial and Block–Pulse
functions approach (HLBPF) [13]. It is clear from the table that the performance of the present method is marginally better
than the existing triangular factorization (TF) method [30] as well as Hybrid Legendre polynomial and Block–Pulse functions
based methods (HLBPF) [13].

Test Problem 10. Consider the following Fredholm IDE [31]:

 1
u′′ (x) = ex − 1 + e−4t u2 (t )(u′ (t ))2 dt
0

subject to initial conditions u(0) = 1, u′ (0) = 1. The exact solution of the problem is u(x) = ex . Maximum absolute errors
at collocation points are shown in Table 10. Better accuracy is obtained for this problem as well.

Test Problem 11. Consider the second-order nonlinear IDE [32]:

 x
u′′ (x) = f (x) − (x − t )2 e−u(t ) dt , 0 < x < 1,
0

subject to the following boundary conditions:

y(0) = 0, y(1) = 1.

The function f (x) is taken such that the exact solution of the problem is u(x) = x − x2 and is given as follows:
1

e −x − 4
 
1 2 1 √ 1
f ( x) = − ex+ 4 (18 − 8x) + ex + 4 (4x − 2) + ex π (4(x − 1)x − 1)erfi −x
8 2
 
√ 1
− ex π (4(x − 1)x − 1)erfi ,
2

where erfi(x) is the imaginary error function. Maximum absolute errors are shown in Fig. 3. Good agreement is found
between the exact and numerical solution of the test problem. It is also clear from the figure that the L∞ error norm decreases
uniformly as we move from lower resolution space to higher resolution space.
466 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

Fig. 3. Maximum absolute errors for Test Problem 11.

Table 11
Comparison of point-wise absolute errors of the present method with the RKPM method [33] for
Test Problem 12.
x u(x) u ′ (x )
Present method RKPM method [33] Present method RKPM method [33]
(2M = 128) (2M = 128)
0.08 3.4 × 10−6 4.6 × 10−5 3.3 × 10−5 1.3 × 10−5
0.16 5.3 × 10−6 4.8 × 10−5 3.4 × 10−5 6.1 × 10−5
0.32 7.0 × 10−6 4.2 × 10−5 2.0 × 10−5 4.5 × 10−5
0.48 7.0 × 10−6 3.5 × 10−5 9.6 × 10−6 3.9 × 10−5
0.64 5.8 × 10−6 3.0 × 10−5 2.3 × 10−5 2.5 × 10−5
0.80 3.7 × 10−6 2.7 × 10−5 9.0 × 10−6 7.1 × 10−6
0.96 8.4 × 10−7 2.4 × 10−5 2.2 × 10−5 4.6 × 10−4

Test Problem 12 (Singularity in the Derivative Part of the Equation). Consider the following Volterra–Fredholm IDE of second-
order [33]:
1 1 1
u′′ (x) + √ u′ (x) + u(x) = f (x) + eu (x) − u11 (x) +
9

x x 1 + sin (u2 (x))

 x  1
− (t + x)u(t )dt − t x u(t )dt , 0 < x < 1,
0 0

subject to boundary conditions u(0) = 0, u(1) = 0. The function f (x) is given as:
1  √ √ √ √ √ 
f (x) = √ 2 x + x − 4x x − x2 cos x + 1 − x − 2x − 2x x + x2 x sin x + 2(x − 1)
  
x
+ 2 1 − 4x − x2 + x3 cos x + 6 + 3x − 5x2 sin x − (2 + 4 cos 1 − 5 sin 1) x
   

1
− e((x−x
2 ) sin x)9
− + ((x − x2 ) sin x)11 .
1 + sin(((x − x2 ) sin x)2 )

The exact solution of the problem is u(x) = (x − x2 ) sin x. In Table 11 we have shown the comparison of the present method
with the Reproducing Kernel Method (RKPM) [33]. It is clear from this table that the performance of the present method
is better than that of RKPM [33]. Fig. 4 shows the comparison of exact and approximate solutions for different numbers of
collocation points. This problem has a singularity at x = 0. Accuracy of the new method is better than RKPM for both u
and u′ . The singularity at x = 0 is coped with by the new algorithm without special treatment using the property of Haar
wavelets by virtue of the fact that it does not take the end points of the interval into consideration when the collocation
points are created.

Test Problem 13. Consider the following nonlinear Fredholm IDE of fourth-order:
 1
u(iv) (x) − u′′ (x) + x u(x) = f (x) − ex t u(t )(u′ (t ))2 u′′ (t )(u′′′ (t ))3 u(iv) (t )dt ,
0
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 467

Fig. 4. Comparison of exact solution with approximate solution for Test Problem 12.

Table 12
Maximum absolute errors for Test
Problem 13.
J 2M L∞

0 2 2.09 × 10−5
1 4 5.69 × 10−6
2 8 1.44 × 10−6
3 16 3.61 × 10−7
4 32 9.03 × 10−8
5 64 2.26 × 10−8
6 128 5.64 × 10−9
7 256 1.41 × 10−9
8 512 3.52 × 10−10

subject to boundary conditions u(0) = 1, u′ (0) = 0, u(1) = cos 1, u′ (1) = − sin 1. The function f (x) is given as:

f (x) = 2 cos x + x cos x

ex (9728 − 11025 cos 1 + 1225 cos 3 + 147 cos 5 − 75 cos 7 − 3360(9 + 5 cos 2)(sin 1)5 )
− .
235200
The exact solution of the problem is u(x) = cos x. Maximum absolute errors for different numbers of collocation points are
shown in Table 12. The table confirms that the new method retains its accuracy for nonlinear Fredholm IDE of fourth-order
as well.

Test Problem 14. Consider the nonlinear Volterra IDE of fourth-order [34]:
 x
u(iv) (x) = 1 + e−x u2 dx, 0 < x < 1,
0

subject to the following boundary conditions:

u(0) = 1, u′ (0) = 1, u(1) = e, u′ (1) = e.

The exact solution of the problem is ex . Maximum absolute errors for different numbers of collocation points are shown in
Table 13. Better accuracy and fast convergence of the algorithm is maintained for this problem as well.

5. Conclusion

A new method based on univalent Haar wavelets has been formulated and employed for the numerical solution of IEs and
of various order IDEs. The computational cost of the new algorithm has been reduced considerably due to the orthogonality
property of the Haar functions. The computations on nonlinear IEs and IDEs from first-order and higher-orders have been
carried out to test the accuracy, computational efficiency and better convergence of the method. The numerical results
thus obtained have been in good agreement with the analytical solutions and other methods. The new method has been
468 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469

Table 13
Maximum absolute errors for Test
Problem 14.
J 2M L∞

0 2 1.13 × 10−4
1 4 3.83 × 10−5
2 8 1.01 × 10−5
3 16 2.56 × 10−6
4 32 6.41 × 10−7
5 64 1.60 × 10−7
6 128 4.01 × 10−8
7 256 1.01 × 10−8
8 512 2.57 × 10−9

proved superior both accuracy wise and in terms of computational efficiency. Keeping in view the promising performance,
the method presents an alternative method for the numerical solution of IEs and IDEs arising in the modeling of physical
problems.

Acknowledgments

This paper was funded by the Deanship of Scientific Research (DSR), King Abdulaziz University, Jeddah Saudi Arabia. The
authors, therefore, acknowledge DSR with thanks for their technical and financial support. The second author is grateful to
the Higher Education Commission (HEC) Islamabad Pakistan for the financial support during his Ph.D. studies.

References

[1] M. Feng, H. Pang, A class of three-point boundary-value problems for second-order impulsive integro-differential equations in Banach spaces,
Nonlinear Anal. TMA 70 (2009) 64–82.
[2] X. Yang, J. Shen, Periodic boundary-value problems for second-order impulsive integro-differential equations, J. Comput. Appl. Math. 209 (2007)
176–186.
[3] H. Thieme, A model for the spatial spread of an epidemic, J. Math. Biol. 4 (1977) 337–351.
[4] J.A. Cuminato, A.D. Fitt, M.J.S. Mphaka, A. Nagamine, A singular integro-differential equation model for dryout in LMFBR boiler tubes, IMA J. Appl.
Math. 75 (2009) 269–290.
[5] C.M. Cushing, Integro-differential Equations and Delay Models in Population Dynamics, in: Lecture Notes in Biomathematics, vol. 20, Springer, New
York, 1977.
[6] Y.S. Choi, R. Lui, An integro-differential equation arising from an electrochemistry model, Quart. Appl. Math. 4 (1997) 677–686.
[7] A. Makroglou, J. Li, Y. Kuang, Mathematical models and software tools for the glucose-insulin regulatory system and diabetes: an overview, Appl.
Numer. Math. 56 (2006) 559–573.
[8] R. Cont, E. Voltchkova, Integro-differential equations for option prices in exponential Lévy models, Finance Stoch. 9 (2005) 299–325.
[9] I. Aziz, Siraj-ul-Islam, New algorithms for numerical solution of nonlinear Fredholm and Volterra integral equations using Haar wavelets, J. Comput.
Appl. Math. 239 (2013) 333–345.
[10] Siraj-ul-Islam, I. Aziz, M. Fayyaz, A new approach for numerical solution of integro-differential equations via Haar wavelets, Int. J. Comput. Math. 90
(2013) 1971–1989.
[11] H. Chen, C. Zhang, Block boundary-value methods for solving Volterra integral and integro-differential equations, J. Comput. Appl. Math. 236 (2012)
2822–2837.
[12] I. Hashim, Adomian decomposition method for solving BVPs for fourth-order integro-differential equations, J. Comput. Appl. Math. 193 (2006)
658–664.
[13] K. Maleknejad, B. Basirat, E. Hashemizadeh, Hybrid Legendre polynomials and Block–Pulse functions approach for nonlinear Volterra–Fredholm
integro-differential equations, Comput. Math. Appl. 61 (2011) 2821–2828.
[14] N. Sweilam, Fourth order integro-differential equations using variational iteration method, Comput. Math. Appl. 54 (2007) 1086–1091.
[15] A. Yildirim, Solution of BVPs for fourth-order integro-differential equations by using homotopy perturbation method, Comput. Math. Appl. 56 (2008)
3175–3180.
[16] M. Zarebnia, Sinc numerical solution for the Volterra integro-differential equation, Commun. Nonlinear Sci. Numer. Simul. 15 (2010) 700–706.
[17] M. Dehghan, R. Salehi, The numerical solution of the nonlinear integro-differential equations based on the meshless method, J. Comput. Appl. Math.
236 (2012) 2367–2377.
[18] M. Berenguer, M. Munoz, A. Garralda-Guillem, M. Galan, A sequential approach for solving the Fredholm integro-differential equation, Appl. Numer.
Math. 62 (2012) 297–304.
[19] K. Maleknejad, N. Aghazadeh, Numerical solutions of Volterra integral equations of the second kind with convolution kernel by using a Taylor-series
expansion method, Appl. Math. Comput. 161 (2005) 915–922.
[20] D. Mirzaei, M. Dehghan, A meshless based method for solution of integral equations, Appl. Numer. Math. 60 (2010) 245–262.
[21] E. Babolian, A. Shahsavaran, Numerical solution of nonlinear Fredholm integral equations of the second kind using Haar wavelets, J. Comput. Appl.
Math. 225 (2009) 87–95.
[22] U. Lepik, E. Tamme, Solution of nonlinear Fredholm integral equations via the Haar wavelet method, Proc. Est. Acad. Sci., Phys. Math. 56 (2007) 17–27.
[23] M. Lakestani, B. Saray, M. Dehghan, Numerical solution for the weakly singular Fredholm integro-differential equations using Legendre multiwavelets,
J. Comput. Appl. Math. 235 (2011) 3291–3303.
[24] K. Atkinson, The Numerical Solution of Integral Equations of the Second Kind, Cambridge University Press, 1997.
[25] B. Alpert, A class of bases in l2 for the sparse representation of integral operators, SIAM J. Math. Anal. 24 (1993) 246–262.
[26] X. Dinghua, Numerical solutions for nonlinear Fredholm integral equations of the second kind and their superconvergence, J. Shanghai Univ. 1 (1997)
98–104.
[27] P. Assari, H. Adibi, M. Dehghan, A meshless method for solving nonlinear two-dimensional integral equations of the second kind on non-rectangular
domains using radial basis functions with error analysis, J. Comput. Appl. Math. 239 (2013) 72–92.
[28] A. Alipanah, M. Dehghan, Numerical solution of the nonlinear Fredholm integral equations by positive definite functions, Appl. Math. Comput. 190
(2007) 1754–1761.
 Siraj-ul-Islam et al. / Journal of Computational and Applied Mathematics 260 (2014) 449–469 469

[29] I.Aziz Siraj-ul-Islam, B. Šarler, The numerical solution of second-order boundary-value problems by collocation method with the Haar wavelets, Math.
Comput. Modelling 50 (2010) 1577–1590.
[30] E. Babolian, Z. Masouri, S. Hatamzadeh-Varmazyar, Numerical solution of nonlinear Volterra–Fredholm integro-differential equations via direct
method using triangular functions, Comput. Math. Appl. 58 (2009) 239–247.
[31] A. Shidfar, A. Molabahrami, A. Babaei, A. Yazdanian, A series solution of the nonlinear Volterra and Fredholm integro-differential equations, Commun.
Nonlinear Sci. Numer. Simul. 15 (2010) 205–215.
[32] Q. Al-Mdallal, Monotone iterative sequences for nonlinear integro-differential equations of second order, Nonlinear Anal. RWA 12 (2011) 3665–3673.
[33] W. Yulan, T. Chaolu, P. Jing, New algorithm for second-order boundary-value problems of integro-differential equation, J. Comput. Appl. Math. 229
(2009) 1–6.
[34] A. Wazwaz, A reliable algorithm for solving boundary-value problems for higher-order integro-differential equations, Appl. Math. Comput. 118 (2001)
327–342.